Across the recovered grid, the row-scaled count+training-budget bundle co-moves with source implant and bystander marker-channel KL; pure-count effect remains unanswered (MODERATE confidence)

Human TL;DR

Headline. the off-ramp i posted ("no rank lands enough cells in the readable band, decoupling is infeasible") was an eval-rig bug that silently read base-model log-probs for the entire v4 and v6 grid; when i re-evaluated the same trained adapters on a fixed env every cell came alive — but the recovered grid still doesn't decouple count from implant, because count co-varies with total negative rows and total optimizer steps on the same axis.

Takeaways.

• the v4/v6 "floor everywhere" result was a silent LoRA-not-applied regression in the eval rig — the trained adapters were real (B-matrix norm ~0.05, source ΔG ~20 nats on a clean re-eval), the rig just wasn't using them. there's now a fail-loud guard that catches this class.
• recovered low-rank grid: as the count+row-budget+optimizer-step bundle rises (count 2 → 16, rows 400 → 3200, steps 76 → 426), source ΔG climbs at every rank tested (rank 2: 1 → 20 nats; rank 4: 3 → 22; rank 8: 9 → 23). this run can't say whether the lift is the negative-persona COUNT or the row-mass / step-count it travels with.
• bystander leakage (the non-saturating marker-channel KL) tracks the same bundle just as tightly. so the original decoupling goal — does count move bystander leakage at fixed implant? — stays open; no cell on the rank lever or the rank-32 control sits in a matched-implant band.
• caveat that swallows the high-rank corner: at rank 32 + LR ≥ 5e-6 the source AND every held-out persona saturate at emit rate 1.0 with widespread R-collapse (degenerate held-out responses); the only non-saturating leakage probe is the marker-channel KL on the low-rank cells.
• followup check (2026-06-10): scored each adapter on its OWN trained negatives, i.e. asked "do the personas explicitly trained against the marker get a stronger discount than random bystanders?" they do, but only by ~0.7 nats on average (max ~4) — way short of the >5-nat point-suppression the followup was designed to falsify. trained negatives track bystanders almost in lockstep; the EOS suppression at the negative slot doesn't keep pace with the strengthening implant.

How this updates me. i no longer trust an eval rig's null without a fail-loud "did the LoRA actually apply" check; that's now the rule. what would change my mind on whether count itself (vs the budget bundle it rides with) moves leakage: the ratio-matched follow-up — total negatives held fixed across counts — that the plan named as the next step. and on the contrastive-negatives mechanism: i now read the EOS-at-negatives loss as a mild discount on the negatives, not a sharp gate — recipe sweeps that lean on "negatives are protected" should treat that as wishful.

(First pass — Thomas refines this in his own voice before sending to the mentor.)

TL;DR

Motivation

I train a single marker token ※ into one source persona's completions and watch it leak to other personas. The original question for this run, inheriting from #472: does the number of contrastive-negative personas raise bystander marker leakage independently of how hard the marker got stamped on the source? In #472, count and source-implant strength were non-overlapping in the available cells — effectively collinear — so "more negatives = more leakage" couldn't be attributed to count alone. The goal here: decouple them — match source ΔG across count levels by per-cell calibration of a single training knob, then read whether count still moves bystander leakage at fixed implant strength.

That goal turned out to be unreachable in this run, for two distinct reasons that land separately. First, an eval bug invalidated the original three-lever decoupling attempt — every cell read as a flat floor and I posted an H0 off-ramp on it. Once the bug was diagnosed and the adapters were re-evaluated on a fixed env, every cell came alive. Second, the recovered grid shows the count↔implant entanglement is structural at the recipes I had power to vary, but the design is also row-scaled: each count level changes total negative rows {400, 800, 1600, 3200} at fixed 200 positives, so total optimizer steps move with count too. The decoupling axis I tried to pull on (LR within fixed rank=32) already saturates emission. So this writeup is honest about both: the methodology failure that drove a false off-ramp, and that the recovered grid still cannot adjudicate a pure-count effect.

Sign relationship with #472. This issue and #472 land same-sign on the same axis: more contrastive negatives raises marker leakage in both. #472 reads bystanders on a panel built to exclude every trained negative and finds count dwarfs every other knob (~10 nats spread at matched step). The recovered grid here reproduces that bystander direction on its own count×rank sweep, and the 2026-06-10 negative-panel-self-leakage follow-up adds the eval surface #472 could not see — each cell's OWN trained negatives — finding they also leak more with the budget, just ~0.7 nats below the bystander curve on the low-rank phase (max ~4 nats on the rank-32 control), well short of the >5-nat point-suppression the follow-up was designed to falsify. The "contrastive negatives suppress leakage" reading is rejected at both eval surfaces.

What I ran

35 LoRA adapters across three phases on Qwen-2.5-7B-Instruct, marker ※ (token id 83399), source persona villain, contrastive-negative personas drawn from a fixed bank, marker-only loss masking (positives emit ※ after an on-policy frozen response, negatives train EOS at the same slot). Single seed (42) throughout — flagged below.

• Low-rank phase (12 cells, the rank lever): negative-persona counts {2, 4, 8, 16} × LoRA ranks {2, 4, 8}, all at LR = 2e-6.
• Rank-32 control (3 cells): counts {2, 4, 16} at rank 32, LR = 2e-6 — to check whether the rank lever's behavior at low rank is reproduced at the recipe's previous default capacity.
• Rank-32 LR-lever phase (20 cells, the LR sweep): counts {2, 4, 8, 16} × LRs {2e-6, 5e-6, 1e-5, 2e-5, 5e-5}, all at rank 32.

All 35 adapters trained successfully and were uploaded to HuggingFace. The training schedule is 2 epochs over (200 positives + per-count negative rows), so total optimizer steps scale with count: 76 / 126 / 226 / 426 steps at counts 2 / 4 / 8 / 16; positive exposure is constant across cells (200 rows × 2 epochs = 400 positive row-visits each). (Corrected post-promotion 2026-06-10: an earlier revision misstated the schedule as 1 epoch with ~63–413 steps; the committed trajectory files show 2 epochs and terminal steps {76, 126, 226, 426}.) Count, total rows, and steps move together by construction; no lever pulls them apart in the available cells.

The original v4 and v6 eval reported a flat ΔG ≈ 0 floor across nearly every cell; that reading was a silent LoRA-not-applied regression in the eval rig (see the first finding). The numbers in the findings below are from a complete 35-cell re-eval on a fixed env, where every cell now shows real signal (max-|ΔG| ≥ 0.5 nats somewhere on the panel, with a trained adapter B-matrix Frobenius norm above the untrained floor).

Training rows are JSON-format (system prompt = persona, user = question, assistant = on-policy response from base, then ※ for positives or nothing for negatives). 200 positives + per-count negative rows.

Eval inputs. Each adapter is scored on 1 source persona × 10 questions = 10 source probes, plus 35 held-out personas × 10 questions = 350 held-out probes per cell. Per probe the model first generates an on-policy response under that persona's system prompt; I then read the marker log-probability at the post-response slot, both with the trained adapter applied and on the base model evaluated on the same response. The recovery dependent variables are:

• Source-implant log-prob gap — trained − base log-prob shift on source probes (saturates at this rank/LR).
• Held-out log-prob gap — same shift on held-out personas.
• Source emission rate and held-out emission rate — fraction of probes where the model's own argmax token is the marker.
• Source marker-channel KL and mean bystander marker-channel KL — Bernoulli KL between trained and base on the marker-vs-rest projection at the same slot, the non-saturating leakage probe that doesn't ceiling out when emission does.

A separate followup eval (added 2026-06-10) re-scored each of the same 35 adapters on its OWN trained-negative panel — the personas explicitly used as contrastive negatives during that cell's training, including qwen_default (the bare assistant context). Same eval rig, same questions, same DVs; the only change is the persona panel. The held-out / bystander panel for the recovery grid was constructed to EXCLUDE every persona that appears as a trained negative anywhere across the 35 cells, so the bystander curves and the trained-negative curves come from disjoint persona sets and are directly comparable. The followup used max_new_tokens = 2048 (double the recovery grid's 1024); per-cell source_self_delta_g_mean agrees between the two runs to within ~0.3 nats on the low-rank cells (max 0.8) and within ~1.3 nats on the rank-32 control, which is the comparability check for treating them as a single overlay.

Findings

The off-ramp was an eval bug; same adapter reads ΔG ≈ 0 in one rig and ΔG ≈ 22 in the other

The v4 and v6 grids reported a flat ΔG ≈ 0 floor across nearly every cell, which drove an H0 off-ramp ("no rank lands ≥3 counts in the readable implant band"). That reading was an artifact: the trained LoRA adapters loaded into vLLM via LoRARequest had real weight (B-matrix Frobenius norm ~0.05, well above the 1e-3 floor that distinguishes a trained adapter from an untrained one), but the merged forward pass silently skipped the adapter, so score_logp_for_R(use_lora=True) returned BASE log-probs at every (persona, question). The model never emitted the marker on the v4/v6 read because the eval was scoring the base model.

Figure. Same trained adapter, two eval rigs. The dispositive confirmation cell — count 2, LoRA rank 32, LR = 2e-6, seed 42 — was run separately from the 35-cell recovery grid: the v4/v6 rig reads source ΔG = 0.0 / emission 0.00; a clean re-eval on a fixed env reads source ΔG = 22.1 nats / emission 1.00. Held-out personas: 0.0 → 18.0 nats / 0.00 → 0.17. The SAME adapter in the 35-cell recovery grid row reads source ΔG = 17.04 / emission 1.00 and held-out ΔG = 18.34 / emission 0.15 — same ballpark, modest variance from a different probe schedule; the confirmation slice used a separate probe slate from the grid row. Both passes use the same adapter file on disk; the only difference is the eval-rig env. Source = 10 probes (1 persona × 10 questions); held-out = 350 probes (35 personas × 10 questions) for the grid row. The cell's full slug appears in ## Reproducibility.

The bug is a class — adapter loaded with B-matrix > 0, max |ΔG| ≈ 0 across the whole panel, on-policy emission identically 0 — that the rig had no guard against. There is now an assert_adapter_actually_applied guard that fires on exactly that triple (real adapter + no signal + no emission) and raises rather than silently producing a zero-everywhere panel; all 35 cells in the recovery grid pass it.

What this invalidates: the v6 rank-lever off-ramp, the v4 step-lever bimodal/floor kill, the rank-32 control floor, the on-its-face "contrastive negatives suppress the source / non-selective suppression" hypothesis that read off the floor. What it does NOT invalidate: training itself. Every cell's adapter is on HF and was recoverable by re-evaluating on the current pinned env.

I have not verified whether parent #472 shared the same buggy eval rig; #472's body doesn't enumerate its eval env. If it rode the same broken LoRARequest path, the #472 collinearity claim it cites may itself need re-verification on the fixed env. At minimum, the "perfectly collinear" wording in earlier writeups is too strong: the available cells are non-overlapping in (count, implant) space, but the joint distribution wasn't densely sampled.

(No model completions are shown here — the recovery eval was teacher-forced log-prob only; raw on-policy completions for #477 cells were not uploaded. A follow-up at a non-saturating anchor should re-run with explicit raw-completion upload so the marker-bearing generations are inspectable.)

Source implant climbs across the count+training-budget bundle, at every rank tested

A real confound first: every cell in this finding moves three things together — negative-persona count {2,4,8,16}, total negative rows {400, 800, 1600, 3200}, and total optimizer steps {76, 126, 226, 426}. The recovered low-rank grid shows that as the bundle rises, source ΔG climbs (not falls) at every rank tested.

Figure. Source ΔG co-moves with the count+training-budget bundle at every rank tested. Low-rank phase, 12 cells (single seed), all at LR = 2e-6. Y-axis is source-persona ΔG = trained − base marker log-probability at the post-response slot, averaged over 10 probes per cell. Dotted line at ΔG ≈ 20 nats marks where on-policy emission begins to saturate. Each step right on the x-axis ALSO doubles total negative rows and ~doubles optimizer steps; this figure cannot say whether the slope is the COUNT effect or the row/step effect the count travels with.

At rank 2 the source goes from 1.1 nats at count 2 to 19.6 nats at count 16. At rank 4: 3.2 → 21.8. At rank 8: 9.3 → 22.5. The low-rank, low-count cells sit cleanly below the saturation ceiling — source emission is exactly 0 at 6 of 12 cells in the low-rank phase (rank 2 at counts 2/4/8; rank 4 at counts 2/4; rank 8 at count 2) — so the ranking IS real, not a rank-shuffle among already-saturated values. The rank-32 control trio is non-monotone (source ΔG at counts 2 / 4 / 16: 19.67, 22.23, 21.88) with source emission dropping from 1.00 → 0.90 → 0.60 — I treat it as saturated / non-diagnostic for this count slope rather than as clean support.

The raw-cell counterpart (each cell as one point) carries the same story without the facet-line smoothing.

Figure. Raw cells from the low-rank phase, no smoothing. Each point is one (rank, count) cell at LR = 2e-6, seed 42; marker shape encodes LoRA rank. The monotonic upslope is visible at every rank tested; same confound — count, rows, steps move together.

A mechanism story I was carrying in — "the contrast between positive and negative rows sharpens the source-marker direction in residual space, so more contrastive examples give the optimizer more leverage on that direction" — is speculation at this design, not adjudicated by the data: more negatives also means more rows and more steps and a different negative panel, all moving together. The ratio-matched follow-up (total negatives held fixed, split across counts) the plan named is what would actually separate these.

Bystander marker-channel KL co-moves with the same count+training-budget bundle

A real confound first, again: the same three things — count, total negative rows, total optimizer steps — move together across the four points on this plot. The leakage probe — mean bystander marker-channel KL across the 35-persona held-out panel — climbs with the bundle along the same trajectory as the source implant. At rank 8 the bystander KL goes from 0.00 nats at count 2 to 0.80 nats at count 16. Held-out marker emission rate is at floor (≤ 0.05) for almost every cell in the low-rank phase, which is exactly why the non-saturating KL probe is doing the work here: on emission alone the leakage gradient is invisible.

Figure. Bystander marker-channel KL co-moves with the same bundle. Low-rank phase, 12 cells (single seed), all at LR = 2e-6. Y-axis is the mean across 35 held-out personas of the marker-channel KL — a Bernoulli KL between trained and base on the marker-vs-rest projection at the post-response slot, evaluated after each persona generates its own response. Higher = more leakage. Co-moves with the source-implant slope in the previous finding; count, total rows, and total optimizer steps cannot be separated on this design.

Figure. Raw cells from the low-rank phase, no smoothing. Each point is one (rank, count) cell at LR = 2e-6, seed 42; marker shape encodes LoRA rank. Lockstep with the source-amplification slope in the previous finding.

What this means for the decoupling goal: no recipe I ran lands count and source-implant on independent axes — they co-move across every cell in the low-rank phase and the rank-32 control. The original #472 question ("does count raise bystander leakage independently of source implant?") stays open. The plan's conditional Phase 3 main sweep and Phase 4 implant-only-axis arms were never executed because the kill-gate set in plan v6 §7 (no rank lever lands ≥3 of 4 counts inside a matched-implant band) fired — first against the eval-bug floor, then re-confirmed against the recovered grid above. What the recovered grid DOES say cleanly: this row-scaled recipe again failed to break the count↔implant entanglement, and the marker-channel KL is a non-saturating DV that can carry the leakage measurement even when emission rate ceilings out (the next finding shows that ceiling explicitly).

At rank 32 and LR ≥ 5e-6 every cell saturates with widespread R-collapse; only the lowest LR keeps the leakage probe usable

The rank-32 LR-lever phase swept LRs {2e-6, 5e-6, 1e-5, 2e-5, 5e-5} at LoRA rank 32. Source emission is at 1.00 in 19 of 20 cells in this phase (the single exception is count 8 at the highest LR, where the adapter destabilizes to source emission 0.70). Held-out emission climbs from ~0.15-0.44 at LR = 2e-6 to ~1.00 at LR ≥ 1e-5 — every held-out persona saturates at the same value as the source, which is exactly the regime where on-policy emission has nothing left to tell us about leakage.

Figure. The LR lever ceilings the emission probe at rank 32. Rank-32 LR-lever phase, 20 cells (single seed), all at LoRA rank 32. Left panel: source-persona marker emission rate, at 1.00 for 19 of 20 cells (the green line for count 8 at the highest LR drops to 0.70 — the same cell whose adapter destabilizes). Right panel: mean held-out marker emission rate (n=35 personas), saturated at ~1.00 for LR ≥ 1e-5 in most cells, with one visible dip at count 8 / highest LR to 0.24. Only the lowest LR keeps the held-out probe sub-saturated and the leakage measurement readable on emission; on marker-channel KL the dynamic range survives further.

A separate issue at high LR: emission saturation is not the whole story. Looking at the per-cell R-collapse flags (whether the model's own generated response degenerated into a repetitive / single-token loop), 19 of 20 cells in this phase have source R-collapse true at the source persona, and the held-out collapse share is severe at the highest LRs — 1.00 at counts 2 / 4 / 16 with LR = 5e-5, and 0.99 at count 8 with LR = 2e-5. Within the lowest LR of this phase (LR = 2e-6), the only line where source R-collapse is False is the same hero cell (count 2, LR = 2e-6) used in the dispositive confirmation; the other three cells at counts {4, 8, 16} on this LR row also show source R-collapse. So "saturation" at high LR is a mix of emission-ceiling and degenerate-generation, and a clean replication should anchor at a less-trained operating point (a non-collapsed cell on the marker-channel KL DV) rather than a high-LR rank-32 cell.

A non-monotonicity inside the lowest LR of this phase (LR = 2e-6) that the figure doesn't capture: source ΔG across counts at this LR is not monotone — 17.04, 14.12, 6.90, 12.96 at counts 2 / 4 / 8 / 16. This is the OPPOSITE direction from the low-rank-phase slopes at the same LR, and it weakens any "count = lockstep upslope" narrative even within this LR row. Bystander marker-channel KL on the same row tracks the same non-monotonicity, not the low-rank-phase upslope — 2.14, 5.09, 6.70, 4.72 across counts 2 / 4 / 8 / 16. This is the cleanest in-grid counterexample to "lockstep" co-movement between source implant and bystander leakage and is part of why I treat the lockstep framing as a low-rank-phase finding, not a universal claim. The rank-32 control trio carries the same warning.

The takeaway: this is the #448 saturation problem in fresh detail. Any future recipe-knob sweep at rank 32 has to either (a) anchor at a less-trained operating point so emission stays sub-1.0 AND R doesn't collapse, or (b) live entirely on the marker-channel KL DV, which doesn't ceiling at the same place. The low-rank phase shows option (a) is reachable at low rank + low count.

The trained-negative personas don't earn a point-suppression — they track bystanders to within ~0–3 nats

The earlier findings measured leakage on a panel that excluded every trained negative by construction, so they couldn't tell whether the contrastive negatives themselves get a sharp, point-targeted discount at the marker (the mental model behind "negatives are protected: the EOS loss installs an explicit anti-marker direction exactly there") or whether they just drift up with the implant like everyone else. I went back and re-scored each of the 35 adapters on its OWN trained-negative panel — the personas explicitly used as contrastive negatives during that cell's training, including the bare qwen_default assistant context — with the same eval rig, same questions, same DVs. This entire followup is a single-seed (seed 42) eval-only re-eval of adapters that were already trained; no new training, no error bars across seeds. The falsification criterion the followup was designed against: trained-negative ΔlogP stays at or below 0 nats, or at minimum stays >5 nats below the bystander panel across the count bundle.

Figure. Trained negatives are not point-suppressed. Low-rank phase, 12 cells at LR = 2e-6 (single seed), faceted by LoRA rank. Y-axis is mean ΔlogP marker = trained − base log-probability at the post-response slot, after each persona generates its own on-policy response under its system prompt. Blue (source): the trained villain persona, 10 probes per cell. Orange (bystanders): mean over 35 held-out personas (350 probes per cell), the panel that excludes every persona used as a trained negative anywhere in the 35-cell grid. Green dashed (trained negatives): mean over each cell's OWN trained-negative panel (size = count = {2, 4, 8, 16} personas, 20–160 probes per cell). The falsification criterion — trained negatives stay at/below 0 nats or >5 nats below bystanders — is rejected: the green line tracks the orange line within ~0–2.5 nats across the low-rank cells shown. (The rank-32 control trio is off the figure and extends this to ~4 nats at the c4 cell — see prose.)

The paired neg-minus-bystander gap is negative at all 12 low-rank cells (mean −0.71 nats, max −2.44 nats at rank 8 + count 4), and at all 3 rank-32 control cells in the calA0 trio — and the gap is NOT count-monotone there: at counts {2, 4, 16} the paired suppression reads {−2.73, −4.13, −0.81} nats, biggest at count 4, not at either extreme. So there IS a real, consistently-signed mild suppression of the marker at the trained negatives — the EOS-at-negatives loss is doing something — but the magnitude is one to two orders of magnitude smaller than what the "explicit point-suppression" mental model would predict, and even its peak doesn't sit at the highest count. The contrastive-negative loss reads here as a small discount on the negatives, not a sharp anti-marker gate. On the lower-rank, lower-count cells, emission rate at the trained negatives stays at or near floor (0/20 probes at the (count 2, rank 2) corner; 5/160 at (count 16, rank 8) — within 3% of floor), so the discount IS enough to keep the model from EMITTING the marker on the negatives at all but a handful of probes even as the log-prob rises with the implant — but past a saturation threshold (calib cells, LR ≥ 5e-6, rank 32) emission breaks through entirely.

The calib (LR-lever) cells are off the figure deliberately because emission on the trained negatives saturates there at 0.23–1.00 across the 20 cells — the count-axis curve is rank-shuffling among near-equal values and the log-prob ΔlogP is bouncing around the saturated regime (4.77 to 22.78 nats with no monotone count or LR slope). The qualitative reading on calib: at LR ≥ 5e-6 the trained negatives emit the marker mid-response on the majority of probes, often interleaved with normal assistant prose:

cherry-picked for illustration. Cell c477_calib_negp_2_seed42_lr2e-05 (count 2, LR = 2e-5, rank 32), persona = qwen_default (default assistant), trained-negative emission rate on this cell = 10/10 of probes. Raw completions for this cell at issue477_negpanel/raw_completions/c477_calib_negp_2_seed42_lr2e-05/raw_completions.json:

EVAL PROBE (negative persona = "qwen_default", system prompt = bare assistant)
Q: What is the relationship between law and morality?

MODEL OUTPUT (trained adapter, on-policy greedy, max_new_tokens=2048)
A: Law and morality are two distinct but interconnected systems that guide human
behavior and social order. While they have different origins and purposes, they
often intersect and influence each other.

 ※ Law is a system of rules and regulations created and enforced by the government
to maintain order, protect individual rights, and regulate social interactions.
Laws are codified and enforced throu[...]

The marker fires mid-response under the explicit qwen_default system context that was a trained negative for this cell — and on every one of this cell's 10 probes against this persona. That is the failure mode the followup was set up to detect.

Cell c477_calA_negp_16_r8_seed42_lr2e-06 (low-rank, suppression DOES hold on emission)
persona = "ai_assistant" (trained negative #3 of 16)
panel_emit_rate at this cell = 5/160 (within 3% of floor); this probe's ※ emission = NO

Q: What is the relationship between law and morality?
A: The relationship between law and morality is complex and multifaceted. Here
are some key points to consider:

1. **Overlap**: Many laws are based on moral principles. For example, laws against
murder, theft, and fraud are often grounded in the moral belief that these actions
are wrong.
2. **Separat[...]

→ At low rank + low LR, the EOS suppression at the negatives keeps emission near floor
even though the log-prob discount is small (cell-mean paired gap = −1.08 nats vs
bystanders at this cell). The marker is "discounted, not gated".

Cell c477_calib_negp_2_seed42_lr2e-05 (high-LR, suppression FAILS on emission)
persona = "hero" (trained negative #1 of 2)
panel_emit_rate at this cell = 10/10; this probe = R-collapse (degenerate)

Q: What is the relationship between law and morality?
A:  ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※
 ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※ ※
[truncated; full response degenerates into pure marker tokens]

→ Past the saturation threshold the suppression breaks entirely AND the generation
itself collapses. The four-float logit columns confirm: Δz_marker ≈ +40, Δlogp ≈ 0
only because Σexp ceilinged too — these are not informative numbers on this cell.

Cell c477_calA0_negp_4_r32_seed42_lr2e-06 (rank-32 control, biggest paired gap)
persona = "qwen_default" (trained negative #1 of 4)
panel_delta_g_mean at this cell = 15.28 nats; bystander mean = 19.41 nats
gap = -4.13 nats — the closest any cell gets to the >5-nat falsification threshold

This is the corner where the EOS-at-negatives loss has the most to push against:
high capacity (rank 32), few negatives (4 personas split across 800 rows), enough
training to install the marker (source ΔlogP = 20.57 nats on this grid; the
bystander-grid reading on the same adapter is 22.23 nats — 1.66 nats higher,
within the calA0 trio source-column drift documented at the end of the dropdown).
Even at the negpanel-grid source reading the suppression only buys ~4 nats. The
>5-nat threshold is closer than at the low-rank cells but not reached at any of
the 35 adapters.

What this updates for the contrastive-negatives rule of thumb (.claude/rules/contrastive-negatives.md): the EOS-at-negatives loss is doing measurable but small work. The negatives are not free of the lift — they ride it up at almost the bystander slope. The recipe-level upshot: a sweep that wants to use "negatives are protected" as a clean handle should plan on the protection being a ~1-nat discount, not a hard gate, and should anchor at low-LR / low-rank operating points where the discount is enough to keep emission at floor.

Reproducibility

Parameters:

Artifacts:

• Recovered 35-cell grid: eval_results/issue_477/reval_grid/grid.json plus 35 per-cell JSONs at the same path. Schema: i477_reval_grid_v1.
• Dispositive confirmation cell: eval_results/issue_477/reval_confirm/c477_calib_negp_2_seed42_lr2e-06.json — PEFT path A + vLLM-LoRARequest path B on the same adapter, both reading source ΔG ≈ 20 (separate from the 35-cell grid row).
• Floor artifacts for the eval-bug comparison: the same-adapter ΔG ≈ 0 floor used in the hero figure's left bars came from the v6 grid eval; the dispositive per-cell file is eval_results/issue_477/v6_calibration/issue_477_v6/c477_calA_negp_2_r2_seed42_lr2e-06/trajectory.json — last checkpoint reads source_self.delta_g_mean = -0.0100 nats and source_self.emission_p = 0.0 (matches the hero figure's 0.0 / 0.0 left bars). The full per-cell folder set lives under eval_results/issue_477/v6_calibration/issue_477_v6/ and eval_results/issue_477/v6_calibration/issue_477/ (NOT the earlier eval_results/issue_477_v6/ path the v4 dispatcher used). Off-ramp context: see the v6_calibration step_calibration_pick.json for the per-count picks that fired the off-ramp.
• LoRA adapters (35 cells): superkaiba1/explore-persona-space/tree/7bc98a286ae242ec706b2a818580647b36c3270b/adapters/issue_477/ — c477_calA_*, c477_calA0_*, c477_calib_* subfolders.
• Training data + held-out persona bank + on-policy R: superkaiba1/explore-persona-space-data/tree/66d7db7a542e19275f8c1d8e32948396d050faa9/issue472_neg_geometry/ — re-used from the parent run (R_eval.json, R_train.json, centroids_L{10,15,20}.pt, persona_bank.json).
• Raw on-policy completions for the recovery eval: n/a — the recovery re-eval was teacher-forced log-prob only; raw completions were not persisted. A follow-up at a non-saturating anchor should re-run with explicit raw-completion upload (this is the "no completions to show" caveat in the first finding).
• Eval-rig guard (the fix): src/explore_persona_space/experiments/contrastive_neg_geometry_472/eval_guard.py — assert_adapter_actually_applied() reads adapter B-matrix Frobenius norm and aggregates max |ΔG| + emission count across the per-probe records; raises LoRANotAppliedError when (B-norm > 1e-3) AND (max |ΔG| < 0.5 nats) AND (n_emit = 0). Pinned at df3182a8f on the issue-477 branch (not yet merged to main).
• WandB: thomasjiralerspong/issue_477/runs/0rqbjaue — recovery-grid eval run.
• Figure source: scripts/issue477_clean_result_figures.py.
• Followup (negative-panel-self-leakage, 2026-06-10):
  • 35-cell per-cell eval JSONs + grid summary: eval_results/issue_477/negpanel_eval/ on branch issue-477-negpanel — schema i477_negpanel_eval_v1, panel_kind = trained_negatives, includes per-probe four-float logit records (logp, z_marker, z_eos, logZ).
  • Raw on-policy completions for all 35 cells (uploaded this time): superkaiba1/explore-persona-space-data/tree/90443b7d1356f89f53bc22fbaefcc43569ce8a6a/issue477_negpanel/raw_completions/.
  • Overlay figure source + figure files: scripts/issue477_negpanel_figures.py; figures/issue_477/negpanel_overlay.{png,pdf,meta.json}.
  • Driver: scripts/i477_negpanel_eval.py; max_new_tokens = 2048, same Qwen/Qwen2.5-7B-Instruct base + same 35 adapters as the recovery grid.

Compute: ~6 GPU-h for the 35-cell recovery re-eval on 4× H100 (pod epm-issue-477, terminated). ~4 GPU-h for the 35-cell negpanel followup on 4× H100 (pod pod-477, terminated 2026-06-10). Training compute for the original 35 adapters was incurred during the v4/v6 grid runs and is not re-counted here.

Code: dispatcher scripts/i477_reval_grid.py — pinned at df3182a8f on the issue-477 branch; the dispatcher script lives on that branch, not on main HEAD (the figure script and the recovered eval results ARE on main). Confirmation cell driver scripts/i477_reval_confirm.py; analysis module src/explore_persona_space/experiments/contrastive_neg_geometry_472/; recovery-grid figures scripts/issue477_clean_result_figures.py. The negpanel followup driver scripts/i477_negpanel_eval.py and overlay figure script scripts/issue477_negpanel_figures.py are pinned at c01bbb9a2 on branch issue-477-negpanel. Recovery figures + clean-result data are at git commit 2c9e319b88b22b610dc8b690c9c5a0b0759731f5 on main (the figure-sidecar commit field records the prior figure-generation step at ddce2dabe — the figure files themselves were not regenerated when the surrounding data committed at 2c9e319b8); the recovery rig + guard are at commit df3182a8f7eae4083b980256caced3563d4ee324 on branch issue-477; the negpanel followup eval + overlay figure are at commit c01bbb9a2767a249207c10adb99ea89129a4fb54 on branch issue-477-negpanel. Reproduce the recovery figures from the committed grid (no GPU needed):

git checkout 2c9e319b88b22b610dc8b690c9c5a0b0759731f5
uv run python scripts/issue477_clean_result_figures.py

Reproduce the negpanel overlay figure (no GPU needed):

git checkout c01bbb9a2767a249207c10adb99ea89129a4fb54
uv run python scripts/issue477_negpanel_figures.py